Center for Analytical Finance University of California ... · I use Automated Content Analysis, ......
Transcript of Center for Analytical Finance University of California ... · I use Automated Content Analysis, ......
Center for Analytical Finance University of California, Santa Cruz
Working Paper No. 42
FOMC Sentiment Extraction and its Transmission
to Financial Markets
Raul Cruz Tadle
Department of Economics, UC Santa Cruz
May 26, 2017
Abstract
I use Automated Content Analysis, adopted from computational linguistics and political science,
to derive sentiments acquired from Federal Open Market Committee (FOMC) meeting documents.
I assign an index to the minutes in order to determine if the sentiments obtained from the
information therein can be classified as hawkish (analogous to improving economic conditions and
stronger inflationary pressures) or dovish (related to deteriorating economic outlook and subdued
price changes). I compare the sentiments of the discussions in the minutes to the sentiments of
information in corresponding FOMC statements released immediately after the meetings and
calculate the surprise component of the relative sentiments. I then evaluate how this news shock in
the minutes impacts broad equity and real estate investment trust indices, as well as the exchange
rate valuation of different world currencies against the U.S. Dollar. My findings indicate that
financial assets respond to the minutes based on the type of news shock they contain and that
financial markets react more significantly during the FOMC's date-based policy guidance period.
Keywords: Financial Market, Asset Price, Central Bank, Fed, FOMC, Monetary Policy,
Exchange Rates
JEL codes: E43, E44, E52, E58, F31, F42
About CAFIN
The Center for Analytical Finance (CAFIN) includes a global network of researchers whose aim is
to produce cutting edge research with practical applications in the area of finance and financial
markets. CAFIN focuses primarily on three critical areas:
• Market Design
• Systemic Risk
• Financial Access
Seed funding for CAFIN has been provided by Dean Sheldon Kamieniecki of the Division of Social
Sciences at the University of California, Santa Cruz.
1 Introduction
The types of public releases provided by the Federal Reserve (Fed) have evolved over time.
Not only has the Fed publicized the quantitative change of its policy tools, especially the federal
funds target rate, but it has also released qualitative information, such as its meeting documents,
which detail the reasoning behind its policies. In particular, there are two sets of meeting
documents whose releases are widely anticipated due to the regular and timely information they
provide. The first is the meeting statements, which are released at 2 PM Eastern Time on the last
day of scheduled FOMC meeting slots. These provide a brief summary of the information used
in the decision-making process of the Federal Open Market Committee (FOMC), the main Fed
committee that determines monetary policy in the United States. They also contain assessments
of the risks on employment and inflation, the two measures that constitute the FOMC dual
mandate.1 The second document type consists of the meeting minutes, which are released three
weeks after the meetings. The minutes correspond to the information from the same meeting,
but contain more extensive nuances and details about the FOMC’s information.2
The most direct and immediate impact of these Fed releases are assessed through the
reactions of financial assets. These reactions are closely examined because they could trigger
significant changes in the real economy.3 Due to the influence that these markets have on the
economy, the Fed has placed much emphasis on them, as is documented in FOMC discussions.
Financial markets, on the other hand, scrutinize the information in FOMC meeting documents
because they discuss policy decisions as well as economic forecasts, specifically those that FOMC
members monitor closely when deciding the appropriate monetary policy to implement.4 These
forecasts, which signal future policy, have significant influence on expectations about economic
fundamentals. Given that asset pricing depends not only on current but also on the likely
evolution of economic indicators - in addition to interest rates - that react largely to monetary
policy, financial market prices move significantly following changes in both the current and
expected future policies.
1See Bernanke et al. (2004) for more discussions regarding the ‘balance-of-risk’ component of the FOMCstatements.
2Although the minutes are not verbatim records like the transcripts, they provide more detailed summaries ofthe discussions and projections than are those found in the statements.
3Bernanke and Kuttner (2005) add further insight about the importance of financial market reactions anddescribe some of the channels through which these reactions could influence the real economy.
4Romer and Romer (2000) find that the FOMC has superior information that is not publicly accessible.
1
Although FOMC documents matter to financial markets, asset valuations significantly change
after the release of the documents only if the information they contain are new or surprising to
market participants. Previous work has found that unexpected information in FOMC statements
cause significant financial market reactions.5 Since the minutes correspond to the same meetings
as the statements, if the statements fully represent the details of the meetings, then the financial
markets should not be reacting to the minutes releases. This is because the information in the
statements are already accounted for by financial markets and the minutes are, in this case,
redundant. On the other hand, if there is new information in the minutes, then asset prices
would adjust to account for this news. My current work thus examines whether considerable
discrepancies in the corresponding FOMC documents are present and evaluates how they affect
asset prices.
To examine the information component from the two meeting documents and to evaluate
whether significant differences exist between them, I use Automated Content Analysis. I extract
information sentiments that proxy for two sets of information: the FOMC economic outlook
based on the current and projected state of the economy and the committee’s overall policy tilt in
the short and medium term. Following the literature on FOMC document sentiment extraction,
I classify the sentiments as hawkish if they relate to more optimistic economic forecasts and/or
larger foreseen inflationary pressures. Hawkish sentiments indicate a higher propensity for the
FOMC to implement contractionary monetary policy. On the other hand, I categorize the
information that portrays a more negative outlook and/or more subdued inflation indicators
as dovish, which conveys a higher chance of the FOMC conducting expansionary monetary
policy. Using these sentiment categories, I derive the unexpected sentiment component of the
minutes. I then evaluate how the surprise sentiment, which I denote as the news shock, affects
financial market variables, such as U.S. and emerging market equity indices and foreign exchange
valuations against the dollar.
My results indicate that unexpectedly more hawkish (dovish) minutes tend to cause negative
(positive) reactions on financial markets, particularly on equity and Real Estate Investment Trust
(REIT) indices. These imply that markets respond negatively (positively) to the perceived higher
likelihood of contractionary (expansionary) monetary policy. The magnitude of the results is
subdued, however, given that asset prices may have two opposing reactions. Hawkish sentiments,
5Some of the papers discussing the impact of FOMC statements on financial markets are Gurkaynak et al.(2005), Lucca and Trebbi (2011), and Rosa (2011).
2
particularly those that emphasize improving economic conditions, have positive impacts on
financial markets.6 In contrast, these same sentiments also signal impending rate increases,
which have negative effects on asset prices. Therefore, given these two opposite reactions to the
same type of sentiment, the observed results are muted.
To circumvent these counteracting effects, I analyze financial market reactions during the
FOMC’s date-based policy guidance period. This form of guidance removed the uncertainty
about the policy path by committing to a specified policy up to a given date in the future.7,8
Therefore, information in the minutes released during this date-based commitment period serves
primarily as a source of economic and inflation outlook. Once this more specific policy guidance
is implemented, I observe that the aforementioned market indices experience positive (negative)
and significant reactions following a more hawkish (dovish) news shock.
My work thus affirms that FOMC minutes releases have significant impacts on financial
market prices. More importantly, this research is the first to evaluate how asset prices vary
based on the type of surprise information sentiments of the minutes. My findings contribute to
the understanding of how FOMC documents may be used to affect the economy in addition to
conventional policy releases, such as announcements of changes in the Fed Funds rate target.
They also give insight into how documents can be better tailored to conduct more effective
central bank transparency.9 Together with the conventional monetary policy tools, information
contained in FOMC documents have a significant impact that can potentially be used to achieve
the Fed mandates of maximum employment and price stability, while keeping interest rates in
moderate levels in the medium term.
The rest of my paper is structured as follows. Section 2 elaborates on the motivation for
the current work. Section 3 discusses relevant literature. Section 4 describes the method used
to calculate sentiments in FOMC documents. Section 5 describes the financial data. It also
conducts an event study to show stylized facts regarding the changes in the standard deviation
of asset prices following the releases of the minutes. Section 6 examines the effects of the news
component of sentiments in the minutes on various financial market indicators, while section 7
6The impact of hawkish sentiments driven mostly by high inflationary pressures are ambiguous.7In particular, in the August 2011 meeting, the FOMC announced that it will be maintaining its then ongoing
expansionary policy until mid-2013. This was later extended to mid-2015.8See Swanson and Williams (2014) for further discussions on the date-based forward guidance.9See Walsh (2007) for a theoretical discussion of optimal central bank dissemination of information. Also see
Blinder et al. (2008) for a survey covering the evolution of central bank communication and transparency.
3
evaluates the robustness of the results. Finally, section 8 presents concluding remarks.
2 Motivation
Most of the empirical literature examining the effectiveness of monetary policy decisions
evaluate the impact of policy implementation, as reflected by changes in the federal funds rate
target, on treasuries of varying maturities, stock markets, and foreign exchange rates.10 However,
the FOMC also provides qualitative information using regularly released documents. This set
of qualitative information not only includes discussions regarding the type of policy that is
implemented; it also incorporates economic projections available to the FOMC during their
meetings as well as indications about the path that monetary policy may take based on such
projections. The evaluation of these FOMC documents is important to further incorporate more
of the information that market participants consider when adjusting the pricing of their assets.
There are two documents, in particular, that are released regularly and which the markets
consider to be available in a timely manner. Briefly after each regularly-scheduled meeting that
occur about every six weeks, a statement is released (around 2 PM Eastern Time). This set
of documents provides a succinct explanation behind the chosen monetary policy. Three weeks
afterwards, the meeting minutes are distributed. These documents relay more information that
are available to the FOMC members during the meeting, such as the committee members’
own projections and the Board of Governor’s staff projections. They also incorporate details
about the then-current measures of various macroeconomic indicators, including unemployment,
housing starts, and inflation.
Several papers have evaluated the qualitative information contained in FOMC documents.11
These research work have focused mainly on the statements, not only because they are released
in conjunction with the monetary policy ruling, but also because they are easier to evaluate
manually.12 Nonetheless, some work assess the reactions of financial markets to the FOMC
minutes releases, but they are limited since they focus on changes in the standard deviation
of financial asset returns immediately following the releases of minutes and do not indicate
10See Kuttner (2001) and Faust et al. (2007) for some examples of this type of evaluation.11Examples of research papers evaluating the qualitative information in FOMC documents include Rosa (2011),
Cannon (2015), Stekler and Symington (2016), and Jegadeesh and Wu (2017).12These statements are about one page long, on average, while the minutes are roughly eight to ten pages in
length.
4
how such reactions vary based on the perceived policy signals and economic outlook from the
information. In addition, no research to date has simultaneously examined the qualitative
information content of both the statements and minutes while also evaluating the impact of
the unexpected information sentiments of the minutes. Therefore, analyzing the information
content of the two sets of documents in a more transparent and consistent method bridges this
gap in the literature and helps determine how markets respond to information differences.
To conduct this study, I use the Dictionary Method of Automated Content Analysis – a
method that is widely used in both computational linguistics and political science – in order to
evaluate the sentiments of the information in the minutes.13 I conduct the same analysis on the
contents of the statements and calculate the relative sentiments of the two documents. From
this, I obtain the surprise component of the relative sentiments and examine the type of asset
pricing adjustments that they cause.
3 Literature
A significant amount of literature has examined the effect of FOMC releases in terms of the
reactions of financial markets. This is because monetary policy is intended to try to affect the
real economy indirectly through the changes in financial assets.14 Given the immediate response
of financial markets to the surprise changes in policy rates, their reactions have become the
measure by which to evaluate the impact of monetary policy.
In the mid-1990’s when the FOMC steered towards greater transparency, the FOMC began
to announce explicit changes to the fed funds target rate shortly after the meetings. These
releases had a significant influence on the expectations formation of financial market participants.
Documenting the differences in the target rate and its future rate counterpart, Nosal et al.
(2001) and Carlson et al. (2006) reiterated that the difference between the two, especially during
recessionary periods, were smaller after the implementation of explicit announcements of the
federal funds target. They argued that even with this simple amount of guidance, the financial
markets improved in forming expectations regarding the target rate.
Ehrmann and Fratzscher (2007) adds to the discussion regarding expectation formation by
13See Grimmer and Stewart (2013) for more discussions regarding the Dictionary Method and AutomatedContent Analysis, in general.
14Bernanke and Kuttner (2005) offers additional discussions regarding the indirect effect of monetary policythrough financial markets.
5
reiterating that through alterations in overnight interest rates, the FOMC indirectly affects the
long-run interest rate. This logic follows from the idea that the interest rate in the future is
simply reflected by a series of short-term interest rates and that any changes in expectations
of future short-term interest rates can very well affect long-term interest rates. Not only can
FOMC actions have repercussions in the short-term, they may also be able to affect the long-run
fundamentals. Therefore, when a central bank is able to communicate well and be transparent
about its actions, expectations about the future can be anchored by the policy path that it maps
out.
The impact of funds rate target movements is not limited to interest rates. Bernanke and
Kuttner (2005) extends Kuttner’s seminal work using fed funds futures and evaluates the impact
of the unexpected component of the target rate changes on equity prices. They find that the
unexpected component cause a very large reaction from equity markets.
Given that FOMC releases occur on scheduled dates, anticipation may also play a role in how
financial markets react to these releases. Lucca and Moench (2015) document a drift in equity
prices beginning a couple of days prior to the announcements of FOMC decisions. Since this
drift does not seem to exist for other macroeconomic releases, they argue that this movement
in equity prices not only demonstrates the amount of attention that the equity markets place
on FOMC policy, but also of how the expectations of policy, itself, may affect equity market
movements.
In addition, financial market indicators not only react to announced alterations in the policy
target, but may also change depending on the discussions that FOMC members hold regarding
monetary policy. To emphasize this point, Aizenman et al. (2016) examine FOMC member
speeches during the ‘tapering tantrum’ period. They find that during this period, tapering
news, particularly those relayed by the chairmen, have large and adverse impact on exchange
rates. The observed effect is largest among those countries with a combination of low external
debt, current account surpluses, and large amounts of international reserves.
There are also research evaluating the impact of FOMC document releases on financial
markets.15 Kiley (2014) analyzes the movements in long-term rates that is resulting from FOMC
15The assessments in these papers are structured around major changes in FOMC documents. For instance, backin May 1999, the FOMC began to release more elaborate and systematically-released statements that describethe rationale for the target as well as the policy tilt. This was augmented just a few months afterwards byincorporating ”balance of risks.” Another major alteration occurred on the meeting minutes back in December2004 when the release of the minutes, which was previously set at roughly six weeks after each meeting, was
6
statements using the first-principal component of short-term interest rates. He then evaluates
how they impact equity markets before and during the Zero Lower Bound (ZLB) period of the
fed funds rate. He finds that equity prices are much more sensitive to the changes in long-term
rates prior to 2009 compared to the period during the ZLB. Additionally, Gurkaynak et al.
(2005) evaluate the impact of FOMC statements on financial markets. They observe that much
of the reactions of the markets are due to the information about the future path of policy as
conveyed by the statements.
Rosa (2013), on the other hand, explains that the minutes, which are issued three weeks
after the meetings, also have significant effects on financial markets as reflected by large spikes
in the volatility of asset prices. This increase in volatility does not last longer than the end
of the trading day, thereby demonstrating that the financial market is able to adjust its asset
valuations shortly after the minutes are released. Extending on the findings of Rosa (2013),
Apergis (2015) also disentangles the impact of FOMC minutes releases on the prices of several
assets using GARCH volatility modeling. He finds that the reaction of the mean and volatility
of these asset prices are more subdued during the financial crisis.16
Indeed, financial market reactions may depend on the type of information the documents
contain and a nascent literature has turned to using Content Analysis, a common method used
in other academic fields, in order to evaluate the qualitative information of the documents. The
methods implemented by this strand of research can be thought of as being two types: heuristic
and automated. The heuristic implementation simply occurs when an individual or group of
individuals manually analyzes the content of the documents and quantifies the information
content. An example of the heuristic implementation is conducted by Rosa (2011). He examines
the impact of monetary policy statements - in conjunction with policy changes - on exchange
rates. He finds that the statements have large and significant effects on the valuation of the U.S.
dollar against other global currencies.
A criticism regarding heuristic Content Analysis is that it is more sensitive to the bias of the
evaluator(s). The manner of implementing this method may be ambiguous and, at times, even
inconsistent. Lucca and Trebbi (2011) shows a case in point as they conduct both a heuristic
expedited to three weeks. This earlier release schedule assured that the information in the minutes are moretimely and would therefore preserve their significance.
16Jubinski and Tomljanovich (2013) also finds significant reaction on volatility but not on the mean of the assetprices. However, they only utilized data between 2006–2007 while Apergis (2015) uses data from 2005-2011.
7
and an automated evaluation of the statements. They find that the heuristic evaluation does
not provide consistent examination especially when compared to the findings of their method
conducted algorithmically. Hence, much of Content Analysis, especially the one used in the
current project, is conducted in an automated manner. This type of content analysis allows
me to be as consistent and transparent as possible in analyzing the contents of the FOMC
documents.17
Moreover, the FOMC documents not only include the rationale for implemented policy but
also contain what is known as forward guidance (FG), or indications of potential changes in
the policy rule in the near future. Campbell et al. (2012) indicate that FOMC FG has two
components: Odyssean and Delphic. Odyssean FG is the component that specifies what policy
will be taken in the future. It binds the monetary policy body to implement it or risk losing
credibility. On the other hand, Delphic FG indicates the forecasts of future fundamentals that
the FOMC uses in their discussions. They imply the nature of policy the FOMC is more likely
to implement.
The sentiment indices I obtain in my current work are analogous measures of the Delphic
FG expressed in FOMC documents. The indices are calculated based on discussions regarding
forecasts of economic fundamentals as well as inflation levels and can therefore proxy for the
perceived amount of risks imposed on the FOMC’s dual mandate. These sentiments hint at the
evolution of policy that the FOMC is considering based on the projections the members observe.
On the other hand, the date-based FG period, the time when the FOMC commits to
implementing a specified policy for a pre-determined set of dates, is analogous to the Odyssean
FG since it indicates a promise to maintain policy for a specified time. As this FG is implemented,
the sentiments in the FOMC documents, particularly in the minutes, proxies for the insight and
beliefs of FOMC members regarding inflation and the economy. They focus much less on policy
expectations and place more emphasis on shaping the public’s beliefs about economic outlook.
Campbell et al. (2016) add to the discussion by empirically and theoretically analyzing the
impact of the two FG components. They use fed funds futures and examine how the release of
the FOMC’s private information about future economic indicators, as represented by Greenbook
Forecasts, may be influencing the expectations for the future fed funds target rate. They find
17See Antweiler and Frank (2004), Boukus and Rosenberg (2006), Apel and Grimaldi (2014), Grimmer andStewart (2013), El-Shagi and Jung (2015), and Stekler and Symington (2016) for additional examples anddiscussions of the use of Content Analysis.
8
significant responses of the expectations for the policy rate and also observe that most of the
reaction can be attributed to Delphic FG.
However, the work that Campbell et al. (2016) present does not directly use the qualitative
information incorporated in the FOMC statements. Their empirical analysis simply uses the
Greenbook forecasts in lieu of the qualitative information. Although these Greenbook forecasts
may be affecting the decisions of the FOMC, they do not fully represent the information used in
the meetings given that the committee members also incorporate other information, such as their
own forecasts, when making monetary policy decisions.18 The information in the documents are
more representative of those used by the FOMC members when they discuss policy given that
they contain a larger scope of details about meeting materials and discussions. Hence, examining
the qualitative information of these documents provide better and more timely understanding of
the FG that the FOMC members are trying to convey. Analyzing these documents will then lead
to a better measure of the type of reaction that financial markets have based on the qualitative
information they acquire.
4 Sentiment Analysis
The FOMC began to release the meeting minutes three weeks after the scheduled meetings,
beginning with the December 14, 2004 meeting. The FOMC members offer the reasoning, within
those minutes themselves, why they expedited the release from the previous schedule of releasing
the documents after six weeks as follows:
Participants noted that the minutes contained a more complete and nuanced explanation
of the reasons for the Committee’s decisions and view of the risks to the outlook than
was possible in the post-meeting announcement, and their earlier release would help
markets interpret economic developments and predict the course of interest rates.
When they were released after a six-week lag, the minutes were seen by some market participants
as offering information that was already ’stale’ or not that useful. Hence, the expedition of the
minutes release not only highlights the relevance of these documents but also enables the FOMC,
18Moreover, the public is also able to consult the information in FOMC documents when making expectationsabout the funds target rate given that they are available much sooner than the Greenbook forecasts, which arepublicly released after a five year lag. Hence, the actual documents are more consistent with the FG that financialmarket participants are able to observe.
9
themselves, to more effectively guide market expectations.
The impact of the minutes depend on the information that they provide. Since the contents
of the minutes that I analyze in my work are qualitative - or expressed in words - instead of
quantitative, I adopt a method that could quantitatively assess their information. To extract a
measure to proxy for the information in the documents, I utilize Automated Content Analysis.
This type of analysis has had many uses in various fields, such as in political science, computational
linguistics, sociology, and even economics, and is typically used to obtain information in documents,
blogs, speeches, and social media posts, especially tweets.19 There are numerous methods
associated with this type of analysis and the choice depends on the structure of the classification.
4.1 Automated Content Analysis: Dictionary Method
In order to create a measure incorporating both the economic outlook as well as the overall
monetary policy inclinations of the FOMC, I conduct the Dictionary Method of Automated
Content Analysis.20 It allows me to determine not which topics are important, but rather, how
they are conveyed in the FOMC documents. Using the method, I can assess how the details
in the meeting documents portray the general outlook regarding unemployment, production,
inflation, and other economic indicators. It requires, however, that I provide categories with
which to classify the information in the documents.
To be consistent with the ongoing literature examining FOMC documents, I utilize two
pre–specified categorization for the documents I am evaluating. These categories are what I
refer to as sentiments. They describe the general outlook that FOMC members have regarding
both economic and inflation conditions as well as signal FOMC policy inclination. They are
classified as either hawkish or dovish. Hawkish sentiments are those that indicate improving
economic outlook and inflationary pressures. Therefore, hawkish sentiments indirectly signal a
higher likelihood of monetary policy tightening. On the other hand, dovish sentiments tend to
emphasize more details regarding deteriorating economic conditions and subdued price changes,
thereby hinting at a higher probability that monetary policy loosening would occur.
The Dictionary Method also requires lists of keywords and related terms, or what are termed
as ‘dictionaries’. The terms in the ‘dictionaries’ are relevant to the categories that are examined
19See Gorodnichenko and Shapiro (2007) for an application of the method on inflation targeting.20Both of these sets of information are included in the same measure. They are combined because disentangling
them is difficult since they are not independent of one another.
10
and symbolize relevant topics that the FOMC emphasizes in their discussions.21
Using the Dictionary Method allows me to implement my document information analysis
in an automated fashion using Python’s Natural Language Processing capability. With the use
of the automated method, I can consistently evaluate the documents as well as enable future
research to replicate the same methods that I have taken.22 Therefore, my results can easily be
validated.
In order to categorize the minutes, I first collect copies of the minutes from the FOMC website
beginning from the one corresponding to the December 14, 2004 meeting up to the minutes of
the December 15 - 16, 2015 meeting. While examining the minutes and the statements, I have
compiled a set of keywords that pertain to inflation and other indicators of economic outlook.
The list of keywords are shown in Table 1.23 The keywords are very similar to those found in
Jegadeesh and Wu (2017), who utilized Latent Dirichlet Allocation (LDA), a different Automated
Content Analysis method, in order to analyze the topic variations in the minutes.24
To evaluate the FOMC minutes (and their corresponding statements in the next subsection),
I have also created lists of positive and negative terms that are found in the FOMC documents.
The lists are shown in Table 2. Their polarization depends on their use in the English lexicon
and are comparable to the terms used by Lucca and Trebbi (2011). I use these sets of terms to
calculate the overall sentiments of the documents.
The keywords used are then categorized as hawkish or dovish. The hawkish keywords are
those terms that, if associated with positive terms, depict hawkish sentiments. The hawkish
keywords, on the other hand, represent dovish sentiments if they are linked to negative terms.
21The keywords and categories chosen are consistent with previous work on FOMC documents. Since thecategories are pre-established and are well-documented in the literature, the analysis is not subject to overfitting,which is a problem that needs to be addressed when using other methods.
22As Lucca and Trebbi (2011) demonstrate, evaluations of FOMC statements may be prone to the bias ofthe individuals examining the documents. They argue that the heuristic approach is inferior to the automatedapproach that they utilize in their analysis since the automated method significantly reduces the amount ofsubjectivity to which their study is exposed.
23There are noticeably more hawkish than dovish keywords. This does not affect the analysis given that thesekeywords reflect the word choice that the FOMC documents use to convey information. Moreover, the sentimentsare not based solely on these keywords but are determined depending on the context under which these keywordsare used.
24The difference with the approach used by Jegadeesh and Wu (2017) is that they use LDA to distinguish topicsthat are emphasized in the minutes. They do not, however, account for the information similar to those in thestatements. They also do not determine the cumulative price changes of different financial market indicators tothe surprise information sentiments. Finally, they examine the reactions of the S&P 500 from the moment theminutes are released up to 15 minutes ahead and therefore do not account for the traders who read the entireminutes as well as those who may have withheld their trading shortly before the minutes release.
11
In contrast, dovish keywords are those terms which portray dovish sentiments when they are
connected to positive terms while these same keywords depict hawkish sentiments if associated
with negative terms.
To illustrate my keyword categorization, I note that the keyword ‘prices’ is a hawkish keyword
and the term ‘higher’ is a positive term. When taken together, ‘higher prices’ depict a relatively
more hawkish sentiment. This is because with higher observed prices, the risk of inflation
increases, thereby causing the Fed to lean towards setting contractionary monetary policies
- actions that could fend off inflationary risks but which may negatively impact employment
levels. On the other hand, ‘higher unemployment’, in which ‘unemployment’ is a dovish keyword,
portrays a relatively more dovish sentiment. This is because with higher unemployment, the
FOMC observes weakness in the economy and hence will be more likely to implement (or
maintain) expansionary monetary policy - policy decisions meant to reduce unemployment but
which could trigger increases in inflation.
After creating my sets of keywords and polarized terms, I implement my initial sentiment
scoring at the sentence level. I separate out the documents into sentences and take out the
punctuations and capitalizations.25 Eliminating the sentences without any of the keywords
follows since these sentences do not portray any significant information regarding the sentiment
of the documents. Hence, for each document d, I am left with nd sentences.26
Each sentence is given a score as follows. Denoting sentence s with keyword type k as sentd,k,
the sentiment score of the sentence depends on the number of positive terms, p, relative to the
number of negative terms, n. Thus, the sentiment score of sentd,k, score(sentd,k), is given by
score(sentd,k) =
1, if k = hawk & p > n
−1, if k = hawk & p < n
−1, if k = dove & p > n
1, if k = dove & p < n
0, otherwise
(1)
25Some researchers also remove commonly occurring ‘stop words’ and afterwards, stem the words - or stripthe words to their roots - before conducting their analysis. I abstain away from doing these changes since thesemodifications could potentially change the meaning and context of the sentences that I am analyzing.
26Sentences that contain both hawkish and dovish keywords are scored the same as those with only hawkishkeywords given that these sentences portray sentiments in the same manner as sentences with only hawkish keys.
12
Sentence scoring examples following this scheme are shown in Table 3. Example Sentence 1
incorporates hawk keywords and more positive than negative terms. Therefore, it is given a
sentence score of ‘+1’, which indicates a hawkish sentiment for that particular sentence. On
the other hand, Example Sentence 3 also includes a hawk keyword but has more negative terms
than positive. Based on this evaluation, it is given a sentence score of ‘-1’, which implies that
the sentence conveys a dovish sentiment. As illustrated, the sentence sentiment scores are not
only determined by the type of keywords they contain, but are also dependent on the relative
number of positive and negative terms they have.
Moreover, I also account for the negation terms when scoring these sentences.27 When a
positive term is in the proximity (that is, if they occur after three words or less) of a negation
term, then it is counted as a negative term. On the other hand, a negative term is counted as
positive if it immediately follows a negation term.
Finally, I calculate the document sentiment score by aggregating the sentence scores and
dividing them by the number of sentences with keywords as shown by
index(d) = 100 ∗ 1
nd
∑score(sd,k) (2)
This calculation controls for the number of relevant sentences in each document, thereby accounting
for potential increases in length of the FOMC documents over time.28 Accounting for the method
used to calculate the sentiment index, figure 3 shows that a higher index value implies that the
document is more hawkish while the index with a lower value implies that the document is more
dovish.
4.2 Executing the Automated Content Analysis on the Minutes
As an initial step, I remove the section in the minutes that consists of the whole statement
released shortly after the meetings since market participants have already gained access to them
three weeks prior. Therefore, any information that this section contains will already have been
accounted for by the market. Next, I conduct the Automated Content Analysis to each of the
FOMC meeting minutes. The sentiment scores of the minutes are shown in figure 4, in which
the gray bar depicts the Great Recession.
27The negation terms used in this analysis are ‘fail’, ‘less’, ‘never’, ‘no’, ‘not’, ‘opposed’, and ‘unlikely’.28Changing nd to be the overall number of sentences per document does not alter the implications of the results.
13
Prior to the economic downturn, the discussions during the FOMC meetings, as reflected
by the minutes, tended to be more hawkish. The more hawkish sentiments reflected the more
optimistic FOMC view about the economy and the inflationary concerns prevalent during this
period. As the recession approached, the minutes sentiment score declined. Given that inflation
concerns, as reflected by discussions in the documents, were still large in the beginning of
the recession, the more dovish sentiments implied that the FOMC began to observe signs of
impending economic weakness. The perceived weakness continued to worsen until its trough in
December 2008, when the FOMC decided to hold the fed funds rate at a range with zero as its
effective lower bound value.29
In the meetings following the official end of the recession in June 2009, more hawkish
sentiments emerged, partially due to the fact that the FOMC also began to monitor the financial
markets, specifically the equity markets, more closely at this time.30 In conjunction with
the gradual reduction in unemployment, the discussions became more hawkish preceding the
economic turbulence that occurred in Europe in 2011. During the height of the European
sovereign debt crisis, there was a short dip in the hawkishness of the FOMC. Only minimal
repurcussions on the US economy occurred after economic tensions in Europe were observed.
Therefore, the sentiments of the information in the FOMC discussions rebounded and continued
to fluctuate above zero.
To further evaluate the indices corresponding to the minutes, I examine whether the observed
trends are statistically significant. I use the nonparametric computational method from Zeileis
et al. (2003) in order to evaluate whether structural breaks in the sentiment indices occur. The
results, consistent with earlier discussions, are shown in Figure 5. Following the BIC criterion,
I find statistical significance for two structural breaks in the series, denoted by the dotted lines
in the figure. The resulting periods coincide with the hawkish sentiment during the Great
Moderation, the pessimism during the economic downturn, and the more subtle hawkishness
coinciding with the recovery period.
29Stekler and Symington (2016) find that it was not until after the collapse of Lehman Brothers in October2008 when the FOMC fully realized the severity of the economic downturn.
30During this period, the equity markets experienced a significant increase and discussions about them havecontributed significantly to the hawkish sentiments of the minutes.
14
4.3 Comparison of the Sentiments in the Minutes and the Statements
The information in the statements is also crucial in assessing how the financial markets
respond to the minutes. A comparison of the indices will not only demonstrate similar the
trends between the two sets of documents, but will also determine whether large deviations
between them occur. Hence, I conduct the same methodology on the statements and compare
them to those calculated for the minutes. The corresponding sentiment indices are shown in
Figure 6.
Comparing the indices of the statements to those of the minutes, I found large deviations
at different periods. The statements tended to be more hawkish than the minutes before the
recession hit. As the Great Recession occurred, the sentiments of the statements fluctuated
from being more dovish at the beginning of the downturn to becoming more hawkish than the
minutes when some signs of weak recovery were observed. Following the collapse of Lehman
Brothers, the sentiments in the statements dropped significantly (even lower than the minutes
sentiments). When the recession ended, the sentiments in the statements began to inch closer
to the sentiments in the minutes. After 2012, the two sentiment indices have began to move
together more closely.31
In general, the information sentiments in the statements are more volatile than those of the
minutes. This is because the discussions in the statements are concentrated on rationalizing
the policy decisions as well as the implied monetary policy path. The details on the minutes,
on the other hand, include conflicting views of the FOMC members. This inclusion of mixed
perspectives about prices and the overall state of the economy results to more subdued sentiments
from the minutes.
5 Data and Stylized Facts
I begin my empirical analysis by determining whether financial market indicators react to
the releases of the minutes. To do this examination, I conduct an event study using intraday
data. The use of high frequency data is common in the literature examining the effects of
31There are two possible explanations for this. The first is that the FOMC have began to more closely matchthe information in the minutes and the statements. The second, and perhaps more likely explanation, is thatthe increasing length of the statements enabled it to procure additional information that would have only beenavailable in the minutes. Thus, the movements of the sentiment indices based on the two documents reflect thischange.
15
macroeconomic announcements and news on financial markets. This is because it helps isolate
the impact of the examined release from the effects of other macroeconomic information. Therefore,
the use of the intraday data in my current work attribute any reactions of the markets to releases
of the minutes.
5.1 Data
The data I am using consists of the NYSE trade quotes of Exchange-Traded Funds (ETFs)
in the period ranging from December 1, 2004 to April 30, 2014. The ETFs are equities that
closely follow a scaled value of various financial market indices. The ETFs used in this analysis
are the SPY, VNQ, EEM, and EWJ. The SPY, the ETF for the U.S. equity market, closely
tracks the S&P 500 stock index as it is priced based on the same basket of stocks that the S&P
500 is calculated with. Bond et al. (2012) highlight the “... potential real effects of financial
markets that stem from the informational role of market prices.” Bernanke and Kuttner (2005)
add that evaluating the general impact on equities is crucial given that equity pricing reflects
some of the changes in both the cost of capital and the private portfolio valuation. Therefore,
equity valuation is a way to examine the effect of new information from FOMC discussions on
financial markets, and also serve as an indirect measure of the potential impact of monetary
policy documents on the real economy.
The VNQ, on the other hand, is the ETF for the Dow Jones Real Estate Investment Trust
(REIT) Index. It closely follows that MSCI US REIT Index, which represents about 99% of the
US equity REITs.32 Therefore, it proxies for the valuation of real estate stocks. It also acts as
a measure of commercial and residential development activities since the MSCI US REIT Index
represents large companies that develop and manage apartment complexes and business offices.
Given that surprise changes in monetary policy affects real estate valuation, it is plausible that
discussions about these policies also trigger some pricing changes.33 It is thus insightful to
examine how document surprises affect the REITs.
EEM is the ETF that tracks the exposure of MSCI large and mid-sized emerging market
companies.34 The higher the value of EEM is, the larger the weighted investment measure
32It does not, however, include mortgage REIT.33See Iacoviello and Minetti (2008) and Bredin et al. (2007) for discussions about the impact of monetary policy
on REITs and on the housing market.34There are 23 emerging markets involved. 42% of the companies followed by the index are from countries that
16
on the companies from emerging markets are. In addition, EWJ tracks the large and mid-
sized companies that are from Japan. It represents roughly 85% of the Japanese stock market.
These two ETF’s measure part of the the international effects of the sentiments of the FOMC
documents. Given that policy rate decisions and discussions about U.S. economic outlook of
the FOMC have confounding effects on global capital and trade flows, the financial markets in
other countries may also be responding to the FOMC document releases.35 Hence, incorporating
these ETF’s are relevant in order to determine how influential the surprise sentiments from the
document information are to foreign equity markets.
I also complement my ETF data with high frequency foreign exchange (forex) data acquired
from Tick Data.36 I utilize the level 1 Best Bid and Offer (BBO) pricing of four of the largest
world currencies against the dollar, namely U.S. Dollar to Japanese Yen, U.S. Dollar to Swiss
Franc, Great Britain Pound to U.S. Dollar, and Euro to U.S. Dollar. This set of data ranges
from May 7, 2008 to Jan. 12, 2016.
For my analysis, I aggregate the data so that I have high frequency data for equally spaced
time intervals. More specifically, I take the last trade price of each five minute interval of the
ETF data. As for the forex data, I utilize the midpoint of the last tick of each five minute
interval to proxy for the forex value.37
5.2 Preliminary Empirical Methodology using Five-Minute Log Returns
To examine whether the volatility spikes following the release of the minutes, I first calculate
the log returns of each financial market variable using the equation
rT,τ = 100 ∗ log(
PT,τPT−1,τ
)
where PT,τ is the effective asset price level at time period T on day τ . Hence, for the ETF data,
PT,τ is the last trade price of time interval T . As for the forex data, PT,τ is the midpoint of the
last tick in time interval T .
constitute BRICS, i.e. Brazil, Russia, India, China, and South Africa.35See Bruno and Shin (2015) for discussions about the effect of monetary policy on cross-border bank capital
flows.36Rosa (2011) finds significant reactions of exchange rates against the U.S. Dollar following monetary policy
and meeting statement surprises.37Using the median quote of each time interval does not qualitatively affect the results.
17
Financial data that is examined at tick-by-tick frequency is highly influenced by outliers.
These values do not represent the activities occurring in the financial markets since they are
typically caused by erroneous placements of bid and ask prices.38 To address potential outliers
in the data, I remove the returns with values that are greater than the 95th percentile or lower
than the 5th percentile of the corresponding time period. The remaining data, therefore, more
accurately represents ongoing market activity.
Afterwards, I separate the returns on the release days of the minutes from the set of days
in which minutes nor statements are released. The type of day is denoted by DS. Therefore,
rT,τ,DS is the five-minute change in the asset price at the given time period T on day τ of day set
DS. The statement days are excluded from this analysis since many papers have already cited
significant price movements of various financial market variables on these days. This follows
from the fact that monetary policy changes, together with a brief rationalization for such a
change – the statements – are announced on this set of days.39
5.3 Volatility Comparisons
Calculating the standard deviation of rT,τ,DS for each five minute interval T determines
whether large reactions in financial markets occur shortly after the release of the minutes. The
calculation is conducted for each set of days using the formula
σT,DS =
√∑ΞDSτ=1 (rT,τ,DS − rT,DS)2
ΞDS
where rT,DS is the mean of rT,τ,DS for time interval T of all ΞDS days in day set DS.
Following Rosa (2013), I compare the standard deviations of days with releases of the minutes
to those with no FOMC statement or minutes releases. The comparisons are shown in Figures 1
and 2. I find that prior to the release of the minutes at 2PM Eastern Time, the standard
deviation of the five-minute returns of the ETFs tended to move about the same during the
release days of the minutes and the non-release days.
Interestingly, as the minutes are released (depicted by the red vertical line), the volatility of
the five minute returns spikes for all of the financial market variables. This large spike in the
38See Brownlees and Gallo (2006) for more discussions about outliers.39The omission of the days with statement releases does not qualitatively affect the results.
18
standard deviations is not observed on days with no FOMC document releases. The volatility
on release days fluctuates then slowly converges to the corresponding standard deviation of
nonrelease days.40
This finding demonstrates that the financial markets react to the releases of the minutes,
implying that the information they contain also matter to market participants. More importantly,
such a reaction may depend on the type of ‘new’ information in the minutes. Hence, I derive a
measure of the new information in this set of documents before assessing the direction of asset
price movements on days when the documents are released.
6 Empirical Methodology and Results
6.1 Assessing the Relative Sentiments of the Minutes
The rise in the volatility of the financial market variables following minutes releases suggests
that there is some significant informational value in the minutes beyond those already presented
in the statements. This information from the minutes must be differentiated from the statements
in order to isolate the information that cause market participants to reevaluate their positions.
To address this idea, I calculate the difference of the two sentiment indices, or what I refer to
as the relative sentiment (RSt), as
RSt = Mt − St (3)
where Mt is the sentiment score of the minutes and St is the sentiment score of its statement
counterpart.
One concern about this measure is that the sentiment scores of the statements and minutes,
even though they were calculated by the exact same method, may not necessarily be comparable.
Hence, in order to have a more consistent measure of the sentiment scores, I take ZMt and ZSt ,
the standardized values of the sentiment scores of the minutes and statements, respectively, and
40To examine the robustness of these stylized facts, sets of days are randomly selected to determine whethersuch movements on minutes release days can be observed in various sets of non-release days. This is not an issue,however, given that no such reactions are observed in the placebo sets examined.
19
calculate the relative sentiment, RZt, using the equation41
RZt = ZMt − ZSt (4)
Figure 7 shows the values ofRSt andRZt over time. Since the two series have similar movements,
I take equation 4 as the value of relative sentiment of the minutes compared to the statements.42
6.2 Expected Component of the Sentiments
Rosa (2011) points out that macroeconomic fundamentals change gradually and that due
to this consistency, the discussions between meetings resemble one another. Consequently, the
statements and minutes from one meeting to the next tend to portray very similar contents,
thereby making the differences in the information and the respective sentiments to be persistent.43
The information conveyed by the previous relative sentiment that has persisted to the current
set of statements and minutes are already accounted for in the pricing of financial assets and
will therefore not cause any significant reaction from markets. However, there is a part of the
relative sentiments that is unexpected and thus cause the markets to react. To extract this ‘news’
component from the rest of the content of the minutes, I first calculate the expected value of RZt
using the Maximum Likelihood Estimation (MLE) method, which provides a parameterization
of the expected component.
Similar to the construction used in Rosa (2011), I use the MLE method and employ the
forecasting regression specification
RZt = γ0 + γRZt−1RZt−1 + γZStZSt + ξt (5)
where I include the statement sentiment index score since market participants have already
observed the discussions in the statements and its information may affect the expectations of
the participants.
The results are given in Table 5. I find that both the RZt−1 and ZSt are significant predictors
of RZt. Hence, I confirm that the past measure of relative sentiments has a significant amount
41In order to standardize the sentiment scores, I subtract the mean and divide by the standard deviation.42Table 4 shows the summary statistics of RZt.43BIC and AIC model selection confirms the consistency of the relative sentiments for one period.
20
of predictive power even after accounting for the information in the statements.
6.3 Surprise Sentiments of the Information in the Minutes
In order to obtain the surprise sentiments, I difference out the expected component. To do
this, I take the results from section 6.2 and compute the expected value of the relative sentiment
as shown by
Et−1(RZt) = γRZt−1RZt−1
= 0.368 ∗RZt−1
for each FOMC minutes.44 Using this measure of expected relative sentiments, the unexpected
component of the relative sentiments, which I denote as the news shock NSt, is given by
NSt = RZt − Et−1(RZt)
= RZt − 0.368 ∗RZt−1
The descriptive statistics for NSt are shown in Table 6.45 Panel A shows that over the course of
the time period under consideration, the average change in relative hawkishness is negative, but
close to zero in absolute value terms. Panel B and Panel C, moreover, provide the descriptive
statistics for the time periods covered by data type.46
6.4 Basic Regression Specification and Results
To examine the impact of the news shock on financial markets, I use multiple regression
analysis. This enables me to measure the effect of the surprise sentiments on the financial
market variables while controlling for time-specific variations as well as perceptions about the
overall equity market riskiness that may have influence over asset returns. The basic regression
44Including the weighted value of the statement score as part of Et−1(RZt) does not qualitatively change theresults.
45The use of an alternative measure of NSt is explored in section 7.6.46The data for the ETFs cover more of the Great Moderation Period while the data for the ForEx pairs covers
a greater span of the ZLB period.
21
specification is given by
rft = α+ βNSNSt + βV IXV IXt + βY Y + ht
In this specification, rft = 100 ∗ log(P4:00,t
P1:50,t
)is the log percentage return of market indicator f
from 1:50 PM, ten minutes prior to the minutes release, to 4:00, the end of trading day t. This
measure determines the cumulative effect of the information in the minutes on the ETFs and
ForEx pairs examined for the rest of the trading day.47 Hence, rV NQt is the log percentage return
of VNQ, the ETF of REIT, from 1:50 PM to 4:00 PM of day t. V IXt is the log percentage
change of the adjusted closing price of VIX on day t.48,49 It serves as a measure of the perceived
financial market riskiness and therefore have been found to have significant impacts on the
changes in asset prices. Furthermore, Y is a vector of year dummy variables used to implement
year fixed effects while ht is the error term of the specification.
The basic regression results are given by Table 7. Panel A gives the results for ETFs while
Panel B provides the results for the different ForEx pairs with the U.S. dollar. I observe that
the estimated news shock coefficients for the ETFs, aside for the EWJ, have a negative sign
while results for the exchange rates, except for USD to JPY, suggest that the dollar appreciates
following surprisingly more hawkish sentiments.
6.5 Main Regression Specification and Results
The previous results are statistically insignificant. This is perhaps due to the combination
of direct and indirect effects of monetary policy discussions. Although the documents signal
monetary policy through the information regarding economic and inflation outlook, the inclusion
of this information may also trigger revisions to the economic outlook of market participants.
Surprisingly more hawkish minutes may increase the perceived likelihood of future contractionary
policy. Since financial asset valuation depends on expected future rates, the signals about more
likely contractionary policy could have adverse effects on asset prices. The positive news shock
47I focus on these financial variables since they have economically larger results than the expectations for thefederal funds rate. This is because much of the time period under consideration cover the zero lower bound periodof the federal funds rate. The impact on the expectations for the federal funds rate, measured using the fed fundsfutures, are examined in more detail in section 7.1.
48VIX is the implied riskiness of the financial market and is measured using the annualized value of the expectedstandard deviation change in the S&P 500 stock index in the next 30 days.
49Data for the adjusted closing price of VIX is obtained from Yahoo Finance on Feb. 11, 2016.
22
of the minutes, on the other hand, may also be positively reshaping the perceptions of future
economic conditions if driven by positive economic outlook. Consequently, financial market
valuation rises.50 The combination of these contradictory effects may cause the financial markets
to have weaker and statistically insignificant reactions to surprise sentiments in the minutes.
One way to account for these counteracting effects is by holding one of them as fixed. Such
an event occurred in the midst of the recovery period following the Great Recession when the
FOMC decided to change its FG implementation to utilize more definitive FG methods. In
particular, Swanson and Williams (2014) find that August 9, 2011 corresponds to the time when
the FOMC first used date-based FG, which replaced the announcement of keeping the policy
rate at their near-zero levels “... for an extended period” with “... at least through mid-2013.”
This change in FG effectively removed the uncertainty regarding the implied path of monetary
policy, at least for a specified amount of time, by indicating that the FOMC will be holding the
prevailing expansionary policy until the middle of 2013.51
Consistent with the findings of Swanson and Williams (2014), I incorporate the FOMC
meeting date - which serves as the beginning of the date-based FG - to evaluate whether
the unexpected information in the minutes have a different impact after the date-based FG
is implemented. To do so, I include in the regression analysis the indicator variable l2011, which
takes a value of 1 for the period after August 8, 2011, and 0 otherwise. The augmented regression
specification measuring the impact on different market variables is given by
rft = α+ βNSNSt + βl2011∗NSl2011 ∗NSt + βV IXV IXt + βl2011 l2011 + βY Y + φt
where l2011 ∗ NSt is the interaction term that represents the additional reaction to the news
shock during the implementation of the date-based FG. In the specification, φt stands for the
error term. The results for the ETF’s are reported in Table 8.52 I find negative (and statistically
significant for VNQ) coefficient estimates for the news shock. These results indicate that before
the date-based FG was implemented, an unexpected one standard deviation increase in the
50The impact of the perceived future economic conditions is ambiguous when discussions about inflation aredriving a significant portion of the news shock.
51The “... at least through mid-2013” phrase was later modified to “... at least through 2014” and “... at leastthrough mid-2015” in the January 2012 and September 2012 meetings, respectively. The phrase was then revisedto “... a considerable time after the asset purchase program ends...” in the December 2012 meeting. None of thesechanges qualitatively affected the results.
52Changing the cutoff date to 2012 gives very similar results.
23
relative sentiments of the minutes caused a decline in the MSCI REIT Index of approximately
51.5 basis points, after accounting for the daily changes in the VIX as well as year-specific
variations. Therefore, these results suggest that the equity and REIT indices react negatively
when the relative sentiments of the minutes is unexpectedly larger compared to the relative
sentiments of the information from the previous minutes.
Moreover, the coefficient estimates for l2011 ∗NSt are positive and highly significant for SPY
and VNQ. To interpret, considering the daily change in the adjusted closing price of VIX as well
as year-specific changes, these results indicate that an unexpected standard deviation increase
in the news shock during the date-based FG period leads to an increase of the S&P 500 index by
46.9 basis points. These results imply that as the date-based FG is implemented, the subsequent
observed reactions are not only larger, in absolute value terms, but are also the opposite of how
the financial markets tended to react to the news in the minutes.
It is insightful to compare my results to the reactions following surprise changes in the
fed funds rate. I find that during the date-based FG, the S&P 500 broad equity stock index
experiences a 19 basis point increase following an unexpected one standard deviation rise in the
relative sentiments of the minutes. As an indirect comparison, Bernanke and Kuttner (2005)
highlight that broad stock indices rose by 1% point following a surprise 25 basis point cut in
the federal funds rate. Hence, the news shock from the minutes have economically significant
impacts on U.S. broad stock indices. In addition, the magnitude of the reactions of the REIT
index is large. Following an unexpected one standard rise in the relative sentiments, the equity
REIT index rises about 47 basis points. This is significant relative to the results obtained by
Bredin et al. (2007). They find that an 82 basis point increase in REIT index a day after
observing a 25 basis point surprise drop in the fed funds rate.
The findings extend beyond the indices tied to domestic markets. The reactions of EEM (the
ETF for Emerging Market Equities) and EWJ (the ETF for Japanese Equities) react similarly
as the SPY. Hence, these values lend support to the notion that not only do domestic financial
markets react to the surprising information content of FOMC documents, but so do foreign
financial markets.
In addition, Table 9 also shows the results for various foreign currencies against the U.S.
dollar. Except against the Japanese Yen, the coefficient estimates for NSt indicate that before
the date-based FG, there are some statistically insignificant appreciation of the US Dollar
24
following an unexpected increase in the relative sentiments. As the date-based FG occurred,
the Japanese Yen has a statistically significant (at the 5 % level) and positive increase in its
valuation against the U.S. dollar following positive news shocks while the rest of the exchange
rates do not significantly react to the unexpected information sentiments of the minutes.53
6.6 Discussions about the Results
The FOMC uses the statements and minutes to share the information they have during
their meetings and further guide financial market expectations about monetary policy. An
indirect effect of these documents is that they also cause market participants to reevaluate their
forecasts regarding economic indicators, based on both the perceived policy path as well as on
how the FOMC believes the economy and inflation will change. Therefore, discussions regarding
improving economic prospects may be conceived as partly stimulatory, given that they cause
optimism in the markets. On the other hand, they also hint at a higher probability of monetary
policy tightening, which has the opposite effect. Therefore, coefficient estimates for NSt are
statistically insignificant for most of the market indicators examined, although they still suggest
that unexpected hawkish (dovish) sentiments during this relatively more uncertain policy period
cause several of the financial market indicators to contract (expand).
More importantly, the interaction term in the regression circumvents the counteracting effects
of the sentiments. This term represents the change in the impact of the news shock during
the date-based FG period. Since the date-based FG period is a time with policy certainty, the
sentiments obtained from the information in the FOMC documents simply reflect the perceptions
about the economy and inflation and do not provide any significant signals regarding changes
in policy. Therefore, the information from the documents do not result to muted impacts on
financial markets.54
Additionally, as shown in Figure 8, low perceived inflation risk coincides with the timing of
the date-based FG period. In the figure, the inflation rate from the previous year is well below
the target of 2%. As a result, the surprisingly hawkish (dovishness) information sentiments of
53The extensions presented in the next section focuses on the results of the ETFs since they are mostly significantwhile the relevant coefficient estimates for the forex pairs are mostly insignificant.
54Considering the changes in the FOMC forward guidance when the Evans Rule was implemented does not alterthe findings that the news shocks driven by optimistic economic outlook have positive impacts on asset prices. Imaintain the division of the time period as the period prior and during the date-based forward-guidance periodto simplify the discussions. This does not alter the qualitative results given that the Evans Rule simply acts asan alternative policy commitment mechanism of the FOMC.
25
the minutes during this period represent projected improvements (deterioration) of economic
fundamentals and therefore leads to increased optimism (pessimism) in financial markets, both
domestic and abroad.
7 Robustness and Extensions
7.1 Determining whether the Sentiments affect the Fed Funds Futures
The discussions in the minutes related to the evaluations of economic and inflationary
factors for the short and medium term determine the sentiment of the FOMC documents. Such
information causes reactions in the financial markets, as portrayed by movements in stock price
indices, real estate investments, and to a lesser degree, exchange rates. The reactions of many
of these markets are significant during the date-based FG. Although the sentiments measure the
indirect signals about the likely evolution of monetary policy, expectations about the target rate
may also react to them since the FOMC utilizes the information from the economic and inflation
forecasts they gather in order to implement monetary policy. It is plausible that despite the delay
in the release of the minutes, the sentiments from these documents also cause further reaction
to the expectations about the policy rate in addition to those triggered by the statements. To
evaluate this possibility, I use federal funds rate futures, which proxy for market expectations
regarding future federal funds rates for different time periods.
To conduct the analysis, I use the main regression from section 6.5 but take the daily log
percentage change of the funds futures as the dependent variable. I examine the fed funds
futures for different horizons, namely the one-month, three-month, six-month, and twelve-month
horizons. The results are shown in Table 10. I find that for fed funds futures with short horizons
(less than six months), the news shock does not have a statistically significant impact. For the
six-month and twelve-month funds futures, there are significant coefficient estimates for NSt
and l2011 ∗NSt.
Interestingly, the coefficient estimate for NSt is positive while the coefficient estimate for
the interaction term is negative. These imply that before the date-based FG was implemented,
the surprise component of the relative sentiments in the minutes led to a small increase in the
price of the futures contract (lowered the expected funds rate) corresponding to a few quarters
ahead.55 Once the date-based FG was placed in the FOMC releases, minutes with positive news
26
shocks had the opposite effect since the futures contract price declined (which consequently
increased the expected funds rate).
It is worth noting that the economic value of these results are very small, especially during the
date-based FG period, when a surprise one-standard deviation increase in the relative sentiments
of the minutes led to a decrease of 0.6 basis points for the six month ahead horizon and a fall
of about 1.5 basis points for the twelve month horizon. These futures rate movements are
much smaller than the 25 basis point increments that occur when actual policy is implemented.
Therefore, these results simply reflect the perceived higher likelihood of future monetary policy
changes and does not signal definitive changes in the policy rate.
7.2 Distinguishing between the Effects of Hawkish and Dovish Sentiments
The length of the minutes enables them to demonstrate relatively more extensive information
that includes mixed signals about the economy. The committee members have their own and, at
times, varying projections for these indicators and thus, their beliefs about the manner in which
the employment and inflation rates will change in the near future may vary significantly. The
minutes then act as a source of qualitative information, especially regarding FOMC member
disagreements that, if naively combined without further evaluation, may not necessarily portray
an accurate depiction of how sentiments affect financial markets.
A particular method that can be used to assess how varying types of beliefs about the
economy as well as inflation could affect financial markets is differentiating between sentences
that hold hawkish sentiments compared to those that portray dovish sentiments. As discussed
earlier, hawkish sentiments emerge from sentences that portray improving or stronger economic
conditions as well as high inflation whereas dovish sentiments are a result of discussions regarding
weakening economic variables and small projected price changes, typically in the form of low
inflation rates.
To evaluate the hawkish and dovish sentiments of each document, I follow a similar approach
as the sentence sentiment scoring shown in equation 1. The main difference is that I calculate the
aggregate number of hawkish and dovish scores separately before dividing each sentiment type
score by the total number of sentences with keywords in each document. This calculation gives
55Determining why the price of longer-horizon fed funds futures rise following positive news shock is beyondthe scope of this paper.
27
Hawk(d) and Dove(d), the overall concentration of hawkish and dovish sentences, respectively.
Next, I calculate the standardized value of Hawk(d) as ZHt,d and the corresponding z-score
of Dove(d) as ZDt,d. As before, this standardization ensures that the sentiment measures are
comparable. Afterwards, I calculate the difference in the corresponding z-scores of the individual
sentiment types between the minutes and their respective statements. The resulting relative
sentiments are denoted as RZHt and RZDt .
I also need to obtain the expected component of the relative sentiments. To do so, I use the
equation
RZgt = γ0 + γRZgt−1RZgt−1 + γZg
t,SZgt,S + ξt
where g stands for either H for hawk or D for dove. Table 11 gives the results. Similar to
previous findings, the persistence of the relative concentration of hawk and dove sentiments are
high and statistically significant. In particular, the coefficient estimate for RZHt−1 is 0.271 while
the estimate for RZDt−1 is 0.419.
I then calculate NSHt and NSDt , the unexpected component of the relative standardized
concentration of hawkish and dovish sentences, as shown by
NSHt = RZHt − Et−1(RZHt ) = RZHt − 0.271 ∗RZHt−1
NSDt = RZDt − Et−1(RZDt ) = RZDt − 0.419 ∗RZDt−1
Table 12 also shows the descriptive statistics of NSHt and NSDt . The regression specification
in this analysis is very similar to the main regression specification in section 6.5. The modified
regression used is given by
rfa,b,t = ν + βNSHNSHt + βNSDNSDt + βl2011∗NSH l2011 ∗NSHt
+βl2011∗NSD l2011 ∗NSDt + βl2011 l2011 + βV IXV IXt + βY Y + ψt
The results are indicated in Table 13. I observe that NSHt and NSDt have negative and positive
coefficients, respectively. On the other hand, l2011 ∗ NSHt and l2011 ∗ NSDt have the opposite
signs as their indicator counterparts. These findings are consistent with the earlier findings
that the concentration of hawkish sentiments before late 2011 caused a decline in the financial
28
markets while those that occur beginning in August 9, 2011 tend to have positive impacts on
the markets.
In addition, the unexpected relative concentration of hawkish sentiments, in particular, has
a large and statistically significant effect during the date-based FG. This further supports the
claim that during this period, hawkish sentiments, and not the decline in the amount of dovish
discussions, are responsible for the positive reactions of financial markets.
7.3 Examining the Reactions of Financial Markets during the ZLB period
To determine whether the impact of the sentiments are attenuated by the occurrence of the
ZLB, I use the regression specification
rft = α+ βV IXV IXt + βl2011 l2011 + βZLB∗NSZLB ∗NSt + βZLBZLB + βY Y + φt
The difference between this specification and that used in section 6.5 is that I replaced NSt and
l2011 ∗NSt with the indicator variable ZLB and its interaction term with NSt.
My results are given in table 14. For each ETF, there are two columns of results. The
first column corresponds to the results from the original regression specification. On the other
hand, the second indicates the results while using the ZLB indicator and its interaction term.
Interestingly, when I examine the results with the ZLB variable, I find positive and significant
coefficient estimates for EEM and EWJ, the equity index measures for emerging countries and
Japan, respectively, while the estimates for SPY and VNQ are positive but insignificant. These
results suggest that the foreign equity markets respond to the surprise component of the relative
hawkishness acquired from minutes information whereas domestic markets have a significant
response only during the date-based FG.
Given that the large companies from Brazil, Russia, India, China, and South Africa (BRICS)
constitute much of those included in the EEM, a significant portion of the companies covered
by EEM and EWJ are linked directly to trading with the U.S. Therefore, the importance of
FOMC document information regarding the state of U.S. economic recovery may help explain
the significant reactions of international equity markets on surprise relative sentiments. Still,
exchange rates do not have significant reactions to the unexpected component of the sentiments.
Therefore, the trade link may only partially explain the reactions of the international equity
29
indices to the news shock during the ZLB period.
7.4 Evaluating the Robustness of the Forward Guidance Date Cutoff
Examining the sensitivity of the date cutoff is crucial given that much of the impact from
the date-based FG may be due to the first few meetings immediately after it was implemented.
It is then possible that after 2011, the impact of the date-based FG is insignificant. To assess
this possibility, I use the modified regression specification
rft = α+ βNSNSt + βpost2011∗NSpost2011 ∗NSt + βV IXV IXt + βpost2011post2011 + βY Y + φt
where I simply replace the variable l2011 and its corresponding interaction term with post2011
and its interaction variable.56 post2011 is an indicator variable that takes a value of 1 for dates
after 2011, and 0 otherwise.
My results are reported in Table 15. I report the original findings using the main specification
as well as those arising from the use of post2011. I find that the results are insensitive to the
change in the timing of the date-based FG. Hence, the implications regarding the impact of the
unexpected hawkishness of the minutes hold.
7.5 Examining the Returns on Minutes Release Days
Earlier findings evaluated the impact on the returns from the surprise sentiment obtained
from the minutes. It is, however, also informative to examine the differences in the returns
solely on minutes release days since these provide a much clearer depiction of the impact of
the surprise component of the minutes. This is because asset price movements on days without
minutes releases may still occur due to other factors. Isolating these effects could give more
clarity about the impact of the news shock.
Table 16 gives the results using the main regression specification. For each of the ETF data,
there are two columns of results. The first indicates the previous findings when all trading
days are considered whereas the second column includes the results when examining only the
minutes release days. I observe that the results for the interaction term l2011 ∗ NSt are larger
for SPY, VNQ, and EEM when only considering the minutes release days. They are also highly
56Changing this indicator variable to another indicator for September 2011 does not qualitatively alter theresults.
30
significant. In contrast, EWJ has a smaller coefficient estimate for the interaction term when
regressing the returns during minutes release days only but still is statistically insignificant.
In addition, the R2 for each of the regressions examining the release days are much higher
than their counterparts examining all days. This is consistent with the notion that much of the
variability of returns on release days largely depends on the variables examined, especially the
surprise component of such documents. Hence, much of the activity for the rest of the trading
day following the release of the minutes are based on their content.
7.6 Using the Residuals as the Unexpected Component of the Minutes
The unexpected component of the minutes is measured by first taking the value of the
expected component through MLE estimation and then subtracting it from the relative zscore.
However, referring back to Eq. 5, the error term, ξt, is the component of the relative z-score that
is not explained by either its lag value and the z-score of the statement sentiment. Therefore,
ξt can be used as the alternative measure of the unexpected component.
Using ξt, the regression specification is given by
rft = α+ βξξt + βl2011∗ξl2011 ∗ ξt + βV IXV IXt + βl2011 l2011 + βY Y + φt
Table 17 gives the results when evaluating returns on all days as well as on the release days
of the minutes. Similar to earlier findings, there are statistically significant and qualitatively
similar results for the interaction term l2011 ∗ ξt for SPY, VNQ, and EEM. Furthermore, the
value of the coefficient estimate for the interaction term is larger in magnitude when evaluating
only minutes release days compared to the examination incorporating all of the days.
7.7 Bootstrap Standard Errors
One additional issue regarding the use of estimated values as independent variables is that
the standard errors, computed in the multiple regression analysis, may be imprecise. This is
because standard errors are calculations of mistakes of predicted values based on the observed
independent variables. If the independent variables used are estimated, then they may not
perfectly represent the actual values of their underlying variables.
A workaround adjustment for the additional mistakes from using estimated variables is using
31
bootstrap standard errors. The analysis with bootstrap standard errors emerged from Monte
Carlo simulation and repeatedly estimates the model using sampling with replacement such
that the standard errors are adjusted to better reflect the mistakes that the model specification
creates. Focusing on the returns on the release days of the minutes, the results are included in
Table 18.
For each of the evaluated ETF’s, there are two columns of findings. The first column
represents the results originally obtained for days with minutes releases. Hence, this set of
results incorporate the original robust standard errors. The second column of each ETF has the
same coefficient estimate but with larger standard errors obtained using the bootstrap method.
Despite the larger standard errors, the coefficient estimates for the interaction term l2011 ∗NSt
are statistically significant at the 10% level for the SPY and EEM and at the 5% level for the
VNQ. Therefore, the results are robust to the larger standard errors obtained from the bootstrap
method.57
8 Concluding Remarks
Communication through documents has increased its relevance to recent monetary policy-
making as central banks from different parts of the world have explored unconventional ways to
stimulate their economies. These documents are intended to affect the expectations formation
about the path that monetary policy will take. Based on the results of my work, the documents
can be used for more than just implementing communication strategies that affect expectations
about policy. They can also be used to guide the outlook regarding inflation and the economy
as well as to impact financial markets to try to influence the real economy.
As it moved away from a regime of secrecy, the FOMC has increased both the frequency
and length of its communication. In doing so, it has increased the amount of information it
releases to the public, especially in terms of the discussions and forecasts released with the
meeting documents. However, the FOMC’s perspective on these markets, in conjunction with
the overall state of employment and price stability may also be self-fulfilling, especially if not
conveyed with much thought and consistency. The FOMC must continue to be careful about
the way in which they discuss their policy-making process, especially as the committee members
57The findings using the error term ξt as the unexpected component are also robust to the use of bootstrapstandard errors.
32
search for a balance between sustaining the recovery of the U.S. economy and preparing for the
future downturns to come.
My current project creates a testable measure of how to systematically evaluate the FOMC
meeting documents, particularly by examining their impact on financial markets in the U.S. and
in other countries. However, the current analysis on financial markets reactions only observe the
immediate effects of these documents. More work must be done to determine just how influential
these meeting documents are to the global economy, as well as to understand if and how the
benefits from such communication are distributed widely.
References
Aizenman, J., Binici, M., and Hutchisonc, M. M. (2016). The Transmission of Federal Reserve
Tapering News to Emerging Financial Markets. International Journal of Central Banking.
Antweiler, W. and Frank, M. Z. (2004). Is All That Talk Just Noise? The Information Content
of Internet Stock Message Boards. The Journal of Finance, 59(3):1259–1294.
Apel, M. and Grimaldi, M. B. (2014). How Informative Are Central Bank Minutes? Review of
Economics, 65(1):53–76.
Apergis, N. (2015). The Role of FOMC Minutes for US Asset Prices Before and After the 2008
Crisis: Evidence from GARCH Volatility Modeling. The Quarterly Review of Economics and
Finance, 55:100–107.
Bernanke, B., Reinhart, V., and Sack, B. (2004). Monetary Policy Alternatives at the Zero
Bound: An Empirical Assessment. Brookings Papers on Economic Activity, 2004(2):1–100.
Bernanke, B. S. and Kuttner, K. N. (2005). What explains the Stock Market’s Reaction to
Federal Reserve Policy? The Journal of Finance, 60(3):1221–1257.
Blinder, A. S., Ehrmann, M., Fratzscher, M., De Haan, J., and Jansen, D.-J. (2008). Central
Bank Communication and Monetary Policy: A Survey of Theory and Evidence. Journal of
Economic Literature, 46(4):910–945.
Bond, P., Edmans, A., and Goldstein, I. (2012). The Real Effects of Financial Markets. Annual
Review of Financial Economics, 4(1):339–360.
33
Boukus, E. and Rosenberg, J. V. (2006). The Information Content of FOMC Minutes.
Bredin, D., O’Reilly, G., and Stevenson, S. (2007). Monetary Shocks and REIT Returns. The
Journal of Real Estate Finance and Economics, 35(3):315–331.
Brownlees, C. T. and Gallo, G. M. (2006). Financial Econometric Analysis at Ultra-high
Frequency: Data Handling Concerns. Computational Statistics & Data Analysis, 51(4):2232–
2245.
Bruno, V. and Shin, H. S. (2015). Capital Flows and the Risk-taking Channel of Monetary
Policy. Journal of Monetary Economics, 71:119–132.
Campbell, J. R., Evans, C. L., Fisher, J. D., and Justiniano, A. (2012). Macroeconomic Effects
of Federal Reserve Forward Guidance. Brookings Papers on Economic Activity, 2012(1):1–80.
Campbell, J. R., Fisher, J. D., Justiniano, A., and Melosi, L. (2016). Forward Guidance and
Macroeconomic Outcomes since the Financial Crisis. In NBER Macroeconomics Annual 2016,
Volume 31. University of Chicago Press.
Cannon, S. (2015). Sentiment of the FOMC: Unscripted. Economic Review-Federal Reserve
Bank of Kansas City, page 5.
Carlson, J. B., Craig, B., Higgins, P., and Melick, W. R. (2006). Fomc Communications and
the Predictability of Near-term Policy Decisions. Futures, 8:10.
Ehrmann, M. and Fratzscher, M. (2007). Communication by Central Bank Committee Members:
Different Strategies, Same Effectiveness? Journal of Money, Credit and Banking, 39(2-3):509–
541.
El-Shagi, M. and Jung, A. (2015). Have Minutes helped Markets to Predict the MPC’s Monetary
Policy Decisions? European Journal of Political Economy, 39:222–234.
Faust, J., Rogers, J. H., Wang, S.-Y. B., and Wright, J. H. (2007). The High-frequency
Response of Exchange Rates and Interest Rates to Macroeconomic Announcements. Journal
of Monetary Economics, 54(4):1051–1068.
34
Gorodnichenko, Y. and Shapiro, M. D. (2007). Monetary Policy when Potential Output is
Uncertain: Understanding the Growth Gamble of the 1990s. Journal of Monetary Economics,
54(4):1132–1162.
Grimmer, J. and Stewart, B. M. (2013). Text as Data: The Promise and Pitfalls of Automatic
Content Analysis Methods for Political Texts. Political analysis, pages 267–297.
Gurkaynak, R. S., Sack, B., and Swansonc, E. T. (2005). Do Actions Speak Louder Than Words?
The Response of Asset Prices to Monetary Policy Actions and Statements. International
Journal of Central Banking.
Iacoviello, M. and Minetti, R. (2008). The Credit Channel of Monetary Policy: Evidence from
the Housing Market. Journal of Macroeconomics, 30(1):69–96.
Jegadeesh, N. and Wu, D. A. (2017). Deciphering Fedspeak: The Information Content of FOMC
Meetings.
Jubinski, D. and Tomljanovich, M. (2013). Do FOMC Minutes Matter to Markets? an Intraday
Analysis of FOMC Minutes Releases on Individual Equity Volatility and Returns. Review of
Financial Economics, 22(3):86–97.
Kiley, M. T. (2014). The Response of Equity Prices to Movements in Long-Term Interest
Rates Associated with Monetary Policy Statements: Before and After the Zero Lower Bound.
Journal of Money, Credit and Banking, 46(5):1057–1071.
Kuttner, K. N. (2001). Monetary Policy Surprises and Interest Rates: Evidence from the Fed
Funds Futures Market. Journal of Monetary Economics, 47(3):523–544.
Lucca, D. O. and Moench, E. (2015). The Pre-FOMC Announcement Drift. The Journal of
Finance, 70(1):329–371.
Lucca, D. O. and Trebbi, F. (2011). Measuring Central Bank Communication: An Automated
Approach with Application to FOMC Statements.
Nosal, E. et al. (2001). How well does the Federal Funds Futures Rate Predict the Future Federal
Funds Rate? Federal Reserve Bank of Cleveland, Economic Commentary,(October), pages
1–4.
35
Romer, C. D. and Romer, D. H. (2000). Federal Reserve Information and the Behavior of
Interest Rates (digest summary). American Economic Review, 90(3):429–457.
Rosa, C. (2011). The High-Frequency Response of Exchange Rates to Monetary Policy Actions
and Statements. Journal of Banking & Finance, 35(2):478–489.
Rosa, C. (2013). The Financial Market Effect of FOMC Minutes. Federal Reserve Bank of New
York Economic Policy Review, 19(2):67.
Stekler, H. and Symington, H. (2016). Evaluating Qualitative Forecasts: The FOMC Minutes,
2006–2010. International Journal of Forecasting, 32(2):559–570.
Swanson, E. T. and Williams, J. C. (2014). Measuring the Effect of the Zero Lower Bound on
Medium-and Longer-Term Interest Rates. The American Economic Review, 104(10):3154–
3185.
Walsh, C. E. (2007). Optimal Economic Transparency. international Journal of Central Banking,
3(1).
Zeileis, A., Kleiber, C., Kramer, W., and Hornik, K. (2003). Testing and Dating of Structural
Changes in Practice. Computational Statistics & Data Analysis, 44(1):109–123.
36
Figure 1: Standard Deviations of ETFs
.04
.08
.12
.16
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
SPY (S&P 500)
.04
.08
.12
.16
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
VNQ (MSCI U.S. REIT).0
4.0
8.1
2.1
6S
tand
ard
Dev
iatio
n
12:00 1:00 2:00 3:00 4:00Time (PM)
EEM (EME Equities)
.04
.08
.12
.16
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
EWJ (MSCI Japan Index)
FOMC Minutes Days Days w/o Minutes or Statements
Figure 2: Standard Deviations of ForEx Pairs
.015
.02
.025
.03
.035
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
USD-JPY
.015
.02
.025
.03
.035
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
USD-CHF
.015
.02
.025
.03
.035
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
GBP-USD
.015
.02
.025
.03
.035
Sta
ndar
d D
evia
tion
12:00 1:00 2:00 3:00 4:00Time (PM)
EURO-USD
FOMC Minutes Days Days w/o Minutes or Statements
37
Figure 3: Interpreting the Sentiment Scores
Figure 4: Sentiment Scores of the FOMC Minutes
-80
-40
040
80S
core
jan05jan05 jan05 jan07jan05 jan07 jan05 jan07 jan09jan05 jan07 jan09 jan05 jan07 jan09 jan11jan05 jan07 jan09 jan11 jan05 jan07 jan09 jan11 jan13jan05 jan07 jan09 jan11 jan13 jan05 jan07 jan09 jan11 jan13 jan15jan05 jan07 jan09 jan11 jan13 jan15
Date
38
Figure 5: Sentiment Index of FOMC Minutes: Structural Breaks
Date
Sco
re
jan05 jan07 jan09 jan11 jan13 jan15
−80
−40
040
80
Figure 6: Sentiment Scores of the Minutes and Statements
-80
-40
040
80S
core
jan05jan05 jan05 jan07jan05 jan07 jan05 jan07 jan09jan05 jan07 jan09 jan05 jan07 jan09 jan11jan05 jan07 jan09 jan11 jan05 jan07 jan09 jan11 jan13jan05 jan07 jan09 jan11 jan13 jan05 jan07 jan09 jan11 jan13 jan15jan05 jan07 jan09 jan11 jan13 jan15
Date
FOMC Minutes FOMC Statements
39
Figure 7: Relative Sentiment Score vs. Relative Z-score
-80
-40
040
80S
core
jan05jan05 jan05 jan07jan05 jan07 jan05 jan07 jan09jan05 jan07 jan09 jan05 jan07 jan09 jan11jan05 jan07 jan09 jan11 jan05 jan07 jan09 jan11 jan13jan05 jan07 jan09 jan11 jan13 jan05 jan07 jan09 jan11 jan13 jan15jan05 jan07 jan09 jan11 jan13 jan15
Date
Non-standardized Standardized
Note: The standardized measure is scaled by 20 in this figure.
Figure 8: Trimmed Mean PCE Inflation
01
23
4In
flatio
n
jan00jan00 jan00 jan02jan00 jan02 jan00 jan02 jan04jan00 jan02 jan04 jan00 jan02 jan04 jan06jan00 jan02 jan04 jan06 jan00 jan02 jan04 jan06 jan08jan00 jan02 jan04 jan06 jan08 jan00 jan02 jan04 jan06 jan08 jan10jan00 jan02 jan04 jan06 jan08 jan10 jan00 jan02 jan04 jan06 jan08 jan10 jan12jan00 jan02 jan04 jan06 jan08 jan10 jan12 jan00 jan02 jan04 jan06 jan08 jan10 jan12 jan14jan00 jan02 jan04 jan06 jan08 jan10 jan12 jan14 jan00 jan02 jan04 jan06 jan08 jan10 jan12 jan14 jan16
Date
40
Table 1: Keywords by Type
Hawkish Terms
business businesses demand economic
economy employment energy equities
equity expansion financial growth
housing income indicators inflation
inflationary investment investments labor
manufacturing outlook output price
prices production recovery resource
securities slack spending target
toll wage wages
Dovish Terms
accommodation devastation
downturn recession
unemployment
41
Table 2: Polarized Terms
Positive Terms
abating* accelerated add advance advanced
augmented balanced better bolsters boom
booming boost boosted eased elevated
elevating expand expanding expansionary extend
extended fast faster firmer gains
growing heightened high higher improved
improvement improving increase increased increases
increasing more raise rapid rebounded
recovering rise risen rising robust
rose significant solid sooner spike
spikes spiking stable strength strengthen
strengthened strengthens strong stronger supportive
up upside upswing uptick
Negative Terms
adverse back below constrained contract
contracting contraction cooling correction dampen
damping decelerated decline declined declines
declining decrease decreases decreasing deepening
depressed deteriorated deterioration diminished disappointing
dislocation disruptions down downbeat downside
drop dropping ebbed erosion fade
faded fading fall fallen falling
fell insufficient less limit low
lower moderated moderating moderation reduce
reduced reduction reluctant removed restrain
restrained restraining restraint resumption reversed
slack slow slowed slower slowing
slowly sluggish sluggishness slumped soft
softened softening stimulate strained strains
stress subdued tragic turmoil underutilization
volatile vulnerable wary weak weakened
weaker weakness
* The term 'abating' is labeled as positive since it is used to describe the deterioration in labor market conditions.
42
Table 3: Examples of Sentence Scoring
Example Sentence 1:
according to survey information expectations of near term inflation⏟ ℎ𝑎𝑤𝑘 𝑘𝑒𝑦
picked upด𝑝𝑜𝑠
in march
consistent with the increase⏟ 𝑝𝑜𝑠
in energy prices⏟ ℎ𝑎𝑤𝑘 𝑘𝑒𝑦
Source: May 24, 2005 Minutes
Sentence Score: +1 (Hawkish) This adds one to the overall document score.
Example Sentence 2:
initial claims for unemployment⏟ 𝑑𝑜𝑣𝑒 𝑘𝑒𝑦
insurance declined⏟ 𝑛𝑒𝑔
further in recent weeks
Source: Aug. 20, 2014 Minutes
Sentence Score: +1 (Hawkish) This adds one to the overall document score.
Example Sentence 3:
outstanding residential mortgage debt declined⏟ 𝑛𝑒𝑔
further in the third quarter of 2010 reflecting
weak⏟ 𝑛𝑒𝑔
housing⏟ ℎ𝑎𝑤𝑘 𝑘𝑒𝑦
activity and tight lending standards
Source: Feb. 16, 2011 Minutes
Sentence Score: -1 (Dovish) This subtracts one from the overall document score.
43
Table 4: Descriptive Statistics of the Relative Z-score of FOMC Documents
Panel A: All Indices (December 2004 - December 2016 FOMC Meetings)
Statistic
Mean 2.88E-08
Standard Deviation 0.819
Minimum -2.305
Maximum 2.158
Minutes Release Days 89
Panel B: Indices covered by ETF data (December 2004 - March 2014 FOMC Meetings)
Statistic
Mean 0.033
Standard Deviation 0.86
Minimum -2.305
Maximum 2.158
Minutes Release Days 75
Panel C: Indices covered by ForEx data (April 2008 - December 2016 FOMC Meetings)
Statistic
Mean 0.006
Standard Deviation 0.814
Minimum -1.758
Maximum 2.158
Minutes Release Days 63
44
Table 5: MLE Regression Results
Dependent Variable:
Variables
0.368***
(0.09)
-0.296***
(0.073)
constant -0.01
(0.073)
Log-likelihood -91.228
Observations 88
Note: *** indicates significance at the 1% level, ** signifies
significance at the 5% level, and * indicates significance at
the 10% level. The numbers in parentheses are the robust
standard errors. The data period is Jan. 1, 2005 to
Jan. 12, 2016. The regression specification is
𝑅𝑍𝑡−1
𝑍𝑡𝑆
𝑅𝑍𝑡
𝑅𝑍𝑡 = 𝛾0 + 𝛾𝑅𝑍𝑡−1𝑅𝑍𝑡−1 + 𝛾𝑍𝑡𝑆𝑍𝑡𝑆 + 𝜉𝑡
45
Table 6: Descriptive Statistics of the News Shock of the Minutes
Panel A: All Indices (December 2004 - December 2016 FOMC Meetings)
Statistic
Mean -0.008
Standard Deviation 0.748
Minimum -2.407
Maximum 2.178
Minutes Release Days 88
Panel B: Indices covered by ETF data (December 2004 - March 2014 FOMC Meetings)
Statistic
Mean 0.014
Standard Deviation 0.784
Minimum -2.407
Maximum 2.178
Minutes Release Days 74
Panel C: Indices covered by ForEx data (April 2008 - December 2016 FOMC Meetings)
Statistic
Mean -0.006
Standard Deviation 0.725
Minimum -1.424
Maximum 1.647
Minutes Release Days 63
46
Tab
le7:
Basi
cR
egre
ssio
nR
esu
lts
Pan
el A
: B
asic
Reg
ress
ion R
esult
s fo
r E
TF
s
Vari
ab
les
SP
Y (
S&
P 5
00
ET
F)
VN
Q (
RE
IT E
TF
) E
EM
(E
ME
Eq
uit
y E
TF
)E
WJ (
Jap
an
Eq
uit
y E
TF
)
-3.1
-24.4
-4.2
0.1
(7.6
)(2
1.2
)(1
0.4
)(6
.1)
Contr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.2
15
0.1
52
0.1
93
0.1
18
# o
f O
bs
2358
1909
2332
2349
Pan
el B
: B
asic
Reg
ress
ion R
esult
s fo
r F
orE
x
Vari
ab
les
U
SD
- J
ap
an
ese
Yen
U
SD
- S
wis
s F
ran
c G
BP
- U
SD
E
uro
- U
SD
-1.6
3.4
-2.5
-6
(3.7
)(3
.1)
(3.6
)(3
.7)
Contr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.0
37
0.0
08
0.0
31
0.0
24
# o
f O
bs
1934
1934
1934
1934
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d f
or
ET
Fs
is D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4 w
hil
e th
e d
ate
per
iod
fo
r F
orE
x i
s
May
7,
20
08
to
Jan
12
, 2
01
6.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+Β𝑌𝑌+ℎ𝑡
𝑁𝑆 𝑡
47
Tab
le8:
Main
Regre
ssio
nR
esu
lts
for
ET
F’s
Vari
ab
les
SP
Y (
S&
P 5
00
ET
F)
VN
Q (
RE
IT E
TF
) E
EM
(E
ME
Eq
uit
y E
TF
)E
WJ (
Jap
an
Eq
uit
y E
TF
)
-10
-51.5
*-1
4.8
-4
(9.3
)(2
7.6
)(1
2.5
)(7
.2)
29.2
**
98.4
***
43.6
**
15.9
(14.1
)(3
3.1
)(1
8.9
)(1
3.7
)
Contr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.2
16
0.1
54
0.1
94
0.1
18
# o
f O
bs
2358
1909
2332
2349
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑙 2011∗𝑁
𝑆𝑙 2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+𝛽𝑙 2011𝑙 2011+Β𝑌𝑌+ℎ𝑡
48
Tab
le9:
Main
Regre
ssio
nR
esu
lts
for
ForE
x
Vari
ab
les
U
SD
- J
ap
an
ese
Yen
U
SD
- S
wis
s F
ran
c G
BP
- U
SD
E
uro
- U
SD
-5.1
0.0
2-2
.5-5
.1
(5.1
)(3
.5)
(5.3
)(4
.8)
15.1
**
7.9
-0.1
-2.2
(7)
(6.5
)(7
.2)
(7.5
)
Contr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.0
39
0.0
09
0.0
32
0.0
24
# o
f O
bs
1934
1934
1934
1934
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
stan
dar
d e
rro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s M
ay 7
, 2
00
8 t
o J
an 1
2,
20
16
. T
he
regre
ssio
n s
pec
ific
atio
n i
s
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑙 2011∗𝑁
𝑆𝑙 2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+𝛽𝑙 2011𝑙 2011+Β𝑌𝑌+ℎ𝑡
𝑙 2011∗𝑁𝑆 𝑡
49
Tab
le10:
Regre
ssio
nR
esu
lts
for
Fed
Fu
nd
sFu
ture
s
Vari
ab
les
FF
1F
F3
FF
6F
F12
0.4
0.8
1.1
1.5
**
(0.4
)(0
.7)
(0.7
)(0
.7)
-0.4
-1-1
.7**
-3***
(0.4
)(0
.7)
(0.7
)(1
)
Contr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.0
35
0.0
62
0.0
68
0.0
57
# o
f O
bs
2599
2602
260
42607
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
stan
dar
d e
rro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
5.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑙 2011∗𝑁
𝑆l 2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+𝛽𝑙 2011l 2011+Β𝑌𝑌+𝜔𝑡
l 2011∗𝑁𝑆 𝑡
50
Table 11: MLE Regression Results for the Expected Component of RZHt and RZDt
Variables
0.271*** 0.419***
(0.093) (0.084)
-0.352*** -0.403***
(0.084) (0.078)
constant -0.013 0.008
(0.081) (0.078)
Log-likelihood -101.085 -97.291
Observations 88 88
Note: *** indicates significance at the 1% level, ** signifies significance at the 5%
level, and * indicates significance at the 10% level. The numbers in parentheses are
the robust standard errors. The data period is Jan. 1, 2005 to Jan. 12, 2016. The
regression specification is
𝑅𝑍𝑡−1𝑔
𝑍𝑡,𝑆𝑔
𝑹𝒁𝒕𝑯
𝑅𝑍𝑡𝑔= 𝛾0 + 𝛾𝑅𝑍𝑡−1
𝑔 𝑅𝑍𝑡−1𝑔
+ 𝛾𝑍𝑡,𝑆𝑔 𝑍𝑡,𝑆
𝑔+ 𝜉𝑡
𝑹𝒁𝒕𝑫
51
Table 12: Descriptive Statistics of NSHt and NSDt
Panel A: All Indices (December 2004 - December 2016 FOMC Meetings)
Statistic
Mean -0.011 0.004
Standard Deviation 0.846 0.84
Minimum -2.32 -2.297
Maximum 2.473 2.232
Minutes Release Days 88 88
Panel B: Indices covered by ETF data (December 2004 - March 2014 FOMC Meetings)
Statistic
Mean -0.005 -0.031
Standard Deviation 0.889 0.887
Minimum -2.32 -2.297
Maximum 2.473 2.232
Minutes Release Days 74 74
Panel C: Indices covered by ForEx data (April 2008 - December 2016 FOMC Meetings)
Statistic
Mean -0.033 -0.012
Standard Deviation 0.737 0.871
Minimum -1.564 -2.297
Maximum 1.929 2.159
Minutes Release Days 63 63
𝑵𝑺𝒕𝑯
𝑵𝑺𝒕𝑯
𝑵𝑺𝒕𝑯
𝑵𝑺𝒕𝑫
𝑵𝑺𝒕𝑫
𝑵𝑺𝒕𝑫
52
Tab
le13:
Regre
ssio
nR
esu
lts:
Separa
ted
Haw
kan
dD
ove
Senti
ments
Va
ria
ble
sS
PY
(S
&P
500 E
TF
)V
NQ
(R
EIT
ET
F)
EE
M (
EM
E E
qu
ity
ET
F)
EW
J (
Ja
pa
n E
qu
ity E
TF
)
-1.4
-10
-3.3
-2.8
(6.7
)(1
7.9
)(9
)(6
.3)
94
211
.80.9
(11.2
)(3
0.5
)(1
5.2
)(8
.4)
29
.6*
**
52
.8*
*27
.92
9.4
*
(11.5
)(2
4.1
)(1
8.9
)(1
6.1
)
-3.9
-52
.3-1
9.8
9.3
(14.3
)(3
5.1
)(2
0.4
)(1
2.4
)
Co
ntr
ols
YE
SY
ES
YE
SY
ES
Yea
r F
EY
ES
YE
SY
ES
YE
S
R2
0.2
17
0.1
54
0.1
94
0.1
19
# o
f O
bs
2358
19
09
23
32
23
49
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
l 2011∗𝑁𝑆 𝑡𝐻
l 2011∗𝑁𝑆 𝑡𝐷
𝑁𝑆 𝑡𝐷
𝑁𝑆 𝑡𝐻
𝑟 𝑡𝑓=𝑣+𝛽𝑁𝑆𝐻𝑁𝑆 𝑡𝐻+𝛽𝑁𝑆𝐷𝑁𝑆 𝑡𝐷+𝛽𝑙 2011∗𝑁
𝑆𝐻l 2011∗𝑁𝑆 𝑡𝐻+𝛽𝑙 2011∗𝑁
𝑆𝐷l 2011∗𝑁𝑆 𝑡𝐷+𝛽𝑙 2011l 2011+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
𝑡+𝛽𝑌𝑌+𝜓𝑡
53
Tab
le14:
Regre
ssio
nR
esu
lts:
Accou
nti
ng
for
ZL
B
Pan
el A
: Re
gres
sio
n R
esu
lts
for
SPY
and
VN
Q
Va
ria
ble
s
S
PY
(S
&P
500 E
TF
)
VN
Q (
RE
IT E
TF
)
[1]
[2]
[3]
[4]
-10
-16.1
-51.5
*-6
7.4
(9.3
)(1
3.6
)(2
7.6
)(5
0)
29.2
**
98
.4*
**
(14.1
)(3
3.1
)
24
.16
8.2
(15
.8)
(53
.8)
Co
ntr
ols
YE
SY
ES
YE
SY
ES
Year
FE
YE
SY
ES
YE
SY
ES
R2
0.2
16
0.2
16
0.1
54
0.1
56
# o
f O
bs
235
8235
819
09
19
09
Pan
el B
: Re
gres
sio
n R
esu
lts
for
EEM
an
d E
WJ
Va
ria
ble
s
E
EM
(E
ME
Eq
uit
y E
TF
)
E
WJ
(Ja
pa
n E
qu
ity
ET
F)
[1]
[2]
[3]
[4]
-14
.8-2
6.1
-4-1
1.6
(12.5
)(1
7.9
)(7
.2)
(9.3
)
43.6
**
15
.9
(18.9
)(1
3.7
)
40.6
**
21
.6*
(20
.7)
(11
.8)
Co
ntr
ols
YE
SY
ES
YE
SY
ES
Year
FE
YE
SY
ES
YE
SY
ES
R2
0.1
94
0.1
94
0.1
18
0.1
2
# o
f O
bs
233
2233
223
49
23
49
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n w
ith
ZL
B i
s
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽ZLB∗𝑁
𝑆𝑍𝐿𝐵∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+𝛽ZLB𝑍𝐿𝐵+Β𝑌𝑌+𝑗 𝑡
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑍𝐿𝐵∗𝑁𝑆 𝑡
𝑍𝐿𝐵∗𝑁𝑆 𝑡
54
Tab
le15:
Regre
ssio
nR
esu
lts:
Ch
an
gin
gth
eD
ate
Cu
toff
Pan
el A
: Re
gres
sio
n R
esu
lts
for
SPY
and
VN
Q
Va
ria
ble
s
S
PY
(S
&P
500 E
TF
)
VN
Q (
RE
IT E
TF
)
[1]
[2]
[3]
[4]
-10
-9.2
-51
.5*
-50
.3*
(9.3
)(9
.2)
(27
.6)
(26
.5)
29
.2*
*9
8.4
***
(14
.1)
(33
.1)
29
.4*
*1
07
.2*
**
(12
.4)
(29
.8)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.2
16
0.2
16
0.1
54
0.1
54
# o
f O
bs
23
58
23
58
19
09
19
09
Pan
el B
: Re
gres
sio
n R
esu
lts
for
EEM
an
d E
WJ
Va
ria
ble
s
E
EM
(E
ME
Eq
uit
y E
TF
)
E
WJ
(Ja
pa
n E
qu
ity
ET
F)
[1]
[2]
[3]
[4]
-14
.8-1
2.5
-4-2
.6
(12
.5)
(12
.3)
(7.2
)(7
.1)
43
.6*
*1
5.9
(18
.9)
(13
.7)
40
**1
3
(17
.8)
(14
.1)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.1
94
0.1
94
0.1
18
0.1
18
# o
f O
bs
23
32
23
32
23
49
23
49
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n u
sed
is
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑝𝑜𝑠𝑡2011∗𝑁
𝑆𝑝𝑜𝑠𝑡2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
+𝛽𝑝𝑜𝑠𝑡2011𝑝𝑜𝑠𝑡2011+Β𝑌𝑌+𝑗 𝑡
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑝𝑜𝑠𝑡2011∗𝑁𝑆 𝑡
𝑝𝑜𝑠𝑡2011∗𝑁𝑆 𝑡
55
Tab
le16:
Regre
ssio
nR
esu
lts:
Retu
rns
on
Min
ute
sD
ays
Pan
el A
: Re
gres
sio
n R
esu
lts
for
SPY
and
VN
Q
Va
ria
ble
s
S
PY
(S
&P
500 E
TF
)
VN
Q (
RE
IT E
TF
)
[1]
[2]
[3]
[4]
-10
-7.9
-51
.5*
-57
**
(9.3
)(9
.5)
(27
.6)
(27
.6)
29
.2*
*3
5.7
**
98
.4*
**1
19
.4*
*
(14
.1)
(17
.6)
(33
.1)
(48
.9)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.2
16
0.4
93
0.1
54
0.4
15
# o
f O
bs
23
58
74
19
09
61
Pan
el B
: Re
gres
sio
n R
esu
lts
for
EEM
an
d E
WJ
Va
ria
ble
s
E
EM
(E
ME
Eq
uit
y E
TF
)
E
WJ
(Ja
pa
n E
qu
ity
ET
F)
[1]
[2]
[3]
[4]
-14
.8-1
4.2
-4-2
.2
(12
.5)
(12
)(7
.2)
(7.5
)
43
.6*
*4
8.3
*1
5.9
13
.3
(18
.9)
(25
.1)
(13
.7)
(15
.2)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.1
94
0.4
59
0.1
18
0.3
83
# o
f O
bs
23
32
73
23
49
74
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑙 2011∗𝑁
𝑆𝑙 2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
𝑡+𝛽𝑙 2011𝑙 2011+Β𝑌𝑌+𝜙𝑡
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
56
Tab
le17:
Regre
ssio
nR
esu
lts:
Usi
ng
Resi
du
als
as
New
sS
hock
Pan
el A
: Re
gres
sio
n R
esu
lts
for
SPY
and
VN
Q
Va
ria
ble
s
S
PY
(S
&P
500 E
TF
)
VN
Q (
RE
IT E
TF
)
[1]
[2]
[3]
[4]
2.4
-5.4
-53
.8*
-53
.8
(6.8
)(8
.5)
(28
.6)
(32
.5)
52
**7
6**
16
8.9
***
21
3.1
**
(23
.4)
(32
.7)
(45
.1)
(83
.9)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.2
16
0.4
94
0.1
54
0.3
81
# o
f O
bs
23
58
74
19
09
61
Pan
el B
: Re
gres
sio
n R
esu
lts
for
EEM
an
d E
WJ
Va
ria
ble
s
E
EM
(E
ME
Eq
uit
y E
TF
)
E
WJ
(Ja
pa
n E
qu
ity
ET
F)
[1]
[2]
[3]
[4]
0.6
-10
.61
0.1
1.7
(11
.7)
(11
.2)
(8.8
)(8
.4)
80
.6*
*1
06
.6*
*3
3.9
32
(32
.3)
(50
.4)
(28
.9)
(30
.7)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
R2
0.1
94
0.4
61
0.1
19
0.3
86
# o
f O
bs
23
32
73
23
49
74
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
rob
ust
sta
nd
ard
erro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
l 2011∗𝜉 𝑡
𝜉 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝜉𝜉 𝑡+𝛽𝑙 2011∗𝜉𝑙 2011∗𝜉 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
𝑡+𝛽𝑙 2011𝑙 2011+Β𝑌𝑌+𝜙𝑡
l 2011∗𝜉 𝑡
𝜉 𝑡
57
Tab
le18:
Regre
ssio
nR
esu
lts:
Boots
trap
Sta
nd
ard
Err
ors
Pan
el A
: Re
gres
sio
n R
esu
lts
for
SPY
and
VN
Q
Va
ria
ble
s
S
PY
(S
&P
500 E
TF
)
VN
Q (
RE
IT E
TF
)
[1]
[2]
[3]
[4]
-7.9
-7.9
-57
**-5
7*
(9.5
)(1
0.6
)(2
7.6
)(2
9.9
)
35
.7*
*3
5.7
*1
19
.4*
*1
19
.4*
*
(17
.6)
(20
.5)
(48
.9)
(52
.3)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
Rep
licat
ion
sn
/a9
,02
4n
/a8
,92
2
# o
f O
bs
74
74
61
61
Pan
el B
: Re
gres
sio
n R
esu
lts
for
EEM
an
d E
WJ
Va
ria
ble
s
E
EM
(E
ME
Eq
uit
y E
TF
)
E
WJ
(Ja
pa
n E
qu
ity
ET
F)
[1]
[2]
[3]
[4]
-14
.2-1
4.3
-2.2
-2.2
(12
)(1
3.4
)(7
.5)
(8.3
)
48
.3*
48
.3*
13
.31
3.3
(25
.1)
(28
.7)
(15
.2)
(19
.2)
Co
ntr
ols
YES
YES
YES
YES
Year
FE
YES
YES
YES
YES
Rep
licat
ion
sn
/a9
,05
4n
/a8
,98
8
# o
f O
bs
73
73
74
74
No
te:
*** i
nd
icat
es s
ign
ific
ance
at
the
1%
lev
el,
** s
ign
ifie
s si
gn
ific
ance
at
the
5%
lev
el,
and
* i
nd
icat
es s
ign
ific
ance
at
the
10
% l
evel
. T
he
nu
mb
ers
in p
aren
thes
es a
re t
he
stan
dar
d e
rro
rs.
Th
e u
nit
s fo
r co
effi
cien
t es
tim
ates
an
d t
hei
r st
and
ard
dev
iati
on
s ar
e b
asis
po
ints
. T
he
dat
a p
erio
d i
s D
ec.
1,
20
04
to
Ap
ril
30
, 2
01
4.
Th
e re
gre
ssio
n s
pec
ific
atio
n i
s
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
𝑟 𝑡𝑓=𝛼+𝛽𝑁𝑆𝑁𝑆 𝑡+𝛽𝑙 2011∗𝑁
𝑆𝑙 2011∗𝑁𝑆 𝑡+𝛽𝑉𝐼𝑋𝑉𝐼𝑋
𝑡+𝛽𝑙 2011𝑙 2011+Β𝑌𝑌+𝜙𝑡
l 2011∗𝑁𝑆 𝑡
𝑁𝑆 𝑡
58